At a glance
Access: Gemini Robotics 2.5 is rolling out to our robotics partners.
Official announcement: Gemini Robotics 2.5
What was announced
Google introduced Gemini Robotics 2.5 on Nov 09 2026 and calls it its most capable vision-language-action (VLA) model to date. The company says the model increases both the breadth of skills a robot can learn and the speed at which it can acquire them. The official post names three areas of progress: dexterity, reliability and safety. Google says Gemini Robotics 2.5 is rolling out to its robotics partners.
What we know
Dexterity. Google evaluated the model across 200 distinct tasks, such as serving food with tongs, tying a bow on a gift box and pipetting a liquid. It says the model achieves a 30% to 100% higher success rate than previous baselines, depending on the type of task, while completing tasks 20% to 30% faster.
Robot types. According to Google, the model works natively with many robot types. It names Apptronik's Apollo 2 humanoid, which has the five-finger, 22 degree-of-freedom SharpaWave hand, and the Franka Duo platform with two-fingered parallel grippers.
Factory pilot. Google partnered with Apptronik and Mercedes-Benz to run Apollo 2 humanoid robots on active production lines in Kecskemét, Hungary, and Sindelfingen, Germany. In the proof-of-concept, Apollo 2 autonomously performs automotive kitting: sorting individual parts from bulk bins into organized trays in the exact order needed for each vehicle. Google reports a 96% success rate. Its embodied reasoning model, Gemini Robotics ER, keeps track of the workflow, and the Gemini Robotics 2.5 actions model controls the robot's hands and arms. A partner statement in the announcement says the models improved dramatically in just a few weeks of post-training and carried over to Mercedes-Benz production with very little drop-off.
Safety. Google says the model uses inference-time reasoning guidance to steer around physical hazards in real time. In its evaluations, this boosted safe task completion around obstacles by 33% and around nearby people by 43%. Google is also releasing ASIMOV-Jailbreak, an open-source benchmark of hundreds of prompts that disguise unsafe requests across dozens of scenes.
Development. Google credits Gemini 4 Argon with greatly accelerating the work. It says Argon acted as a simulation engine, generating thousands of varied practice runs from a real-world camera scene and building a digital twin.
Official samples and early work
- Official sample: video from the official post
- Official sample: video from the official post (2)
- Official sample: video from the official post (3)
- Official sample: video from the official post (4)
What we do not know yet
- Availability. Google has said only that the model is rolling out to its robotics partners. It has not announced wider access, pricing or a date.
- Baselines. Google has not named the previous baselines behind the 30% to 100% and 20% to 30% figures.
- Pilot detail. Google has not stated the number of runs, the time span or the exact definition of success behind the 96% figure.
- Independent testing. The results are Google's own evaluations. The announcement cites no outside replication.
What it means for agent work
Read 96% as a pilot number. Kitting is a bounded, repetitive task, and on a line where parts must arrive in order the remaining failures matter. Ask any partner what happens on a miss.
Planning and control are separate components. In the Mercedes-Benz pilot, one model tracks the workflow and another controls the hands and arms. For teams designing agent systems, that is a concrete example of keeping planning and execution distinct and inspectable.
Measure safety on its own. Google reports hazard-steering results separately from task success. Any evaluation plan for physical or operational agents should keep those measures apart too.
Plan for adaptation. The few weeks of post-training described for the Mercedes-Benz work suggest that site-specific tuning is part of deployment. Budget for it rather than assuming a model works out of the box.
Most teams cannot build on it yet. With access limited to robotics partners, watch for Google's next word on availability. Archon Agents is a managed AI operations company, and if you are weighing how governed agents fit your own operations, you can Book a Consultation.